|
Michael C. Frank
@
mcxfrank
Palo Alto, CA
|
|
Developmental psychologist at Stanford studying language, thought, babies, pragmatics, numbers, cognition, teaching, and learning. Reader, climber, runner, dad.
|
|
|
3,326
Tweets
|
970
Following
|
3,383
Followers
|
| Tweets |
|
|
Michael C. Frank
@mcxfrank
|
2h |
|
It's taking us 3+ hours to go through a single paragraph of results in detail. I think this is way too much to ask of reviewers!
|
||
|
||
| Michael C. Frank retweeted | ||
|
Russ Poldrack
@russpoldrack
|
23h |
|
great thread! twitter.com/mcxfrank/statu…
|
||
|
|
||
| Michael C. Frank retweeted | ||
|
Tobias Wood
@spinicist
|
23h |
|
Methods sections are basically lossy compression. Very bad lossy compression. twitter.com/mcxfrank/statu…
|
||
|
|
||
|
|
Michael C. Frank
@mcxfrank
|
Jul 16 |
|
Of course robustness still a major issue for us too.
|
||
|
|
||
|
|
Michael C. Frank
@mcxfrank
|
Jul 16 |
|
Oh - so this is already huge progress! We need to take that step.
|
||
|
|
||
|
|
Michael C. Frank
@mcxfrank
|
Jul 16 |
|
This is all small-scale lab expts.
|
||
|
|
||
|
|
Michael C. Frank
@mcxfrank
|
Jul 16 |
|
Very interesting, would love to hear more. My impression is AERA articles typically use pre-existing data, specify regression?
|
||
|
|
||
|
|
Michael C. Frank
@mcxfrank
|
Jul 16 |
|
There is a continuum of standards of course. At a minimum, code provides a better guide than verbal description for what was done.
|
||
|
|
||
|
|
Michael C. Frank
@mcxfrank
|
Jul 16 |
|
RMarkdown renders to word format actually, big win in that case.
|
||
|
|
||
|
|
Michael C. Frank
@mcxfrank
|
Jul 15 |
|
A lot of us felt like, let's deal with reality in our own labs rather than responding to anon ethnography. @hardsci was less dismissive...
|
||
|
|
||
|
|
Michael C. Frank
@mcxfrank
|
Jul 15 |
|
Yes definitely! It helps explain what you did more explicitly.
|
||
|
|
||
|
|
Michael C. Frank
@mcxfrank
|
Jul 15 |
|
And the most important thinks then IMO: 1. Exclusion rule 2. Exact stat test 3. Prediction/ interpretation for test result patterns.
|
||
|
|
||
| Michael C. Frank retweeted | ||
|
Heather Urry
@HeatherUrry
|
Jul 15 |
|
Yes, this! Loving @FrederikAust's papaja. Report numbers in ms *directly from R lists* for error-free, reproducible results. Friggin' magic. twitter.com/mcxfrank/statu…
|
||
|
|
||
|
|
Michael C. Frank
@mcxfrank
|
Jul 15 |
|
ugh, thanks, here it is: babieslearninglanguage.blogspot.com/2015/11/preven…
|
||
|
|
||
|
|
Michael C. Frank
@mcxfrank
|
Jul 15 |
|
11/ In sum: Computational reproducibility is a MAJOR issue. Journal/funder policy must reflect critical importance of code sharing. [end]
|
||
|
|
||
|
|
Michael C. Frank
@mcxfrank
|
Jul 15 |
|
10/ RENDER YOUR PAPER. Copy/paste of stats is VERY error-prone. RMarkdown a good solution. babieslearninglanguage.blogspot.com/201...
|
||
|
|
||
|
|
Michael C. Frank
@mcxfrank
|
Jul 15 |
|
9/ SHARE CODE WITH DATA. Of course, it's better to version control/packrat/dockerize. But ANY code/syntax is WAY better than none.
|
||
|
|
||
|
|
Michael C. Frank
@mcxfrank
|
Jul 15 |
|
8/ 4. ALL of my N=3 papers had >1 issue (typo/ error). In *one paragraph*. No conclusions changed, but something def wrong in each. SO:
|
||
|
|
||
|
|
Michael C. Frank
@mcxfrank
|
Jul 15 |
|
7/ 3. ANOVA is the root of all evil. Every program does repeated measures differently, and specifying ANOVA in prose is VERY imprecise.
|
||
|
|
||
|
|
Michael C. Frank
@mcxfrank
|
Jul 15 |
|
6/ 2. There's LOTS of guesswork. Translating code to prose and back to code is very lossy, and you often have to extrapolate key steps.
|
||
|
|
||